Reviewing code is a skill

Recent discussions in developer circles have produced a range of opinions on the purpose and value of code review, from claims that it is a bottleneck to assertions that LLMs outperform humans. The article positions itself against these extremes by framing review as a skill that can be cultivated, rather than a static process or a purely automated task. This reframing shifts the focus from debating the existence of review to improving its execution.

Empirical studies cited in the post reveal that developers expect reviews to serve multiple roles: teaching, reinforcing coding standards, controlling changes, and preventing defects. While defect detection remains a primary driver, the studies also note that reviews foster knowledge transfer and broader team awareness, benefits that are not captured by many current review tools. Recognizing these multiple objectives helps teams design review guidelines that align with broader engineering goals.

The author’s central claim is that reviewers can become better through practice and instruction, and that the upper limits of review performance are not yet known. This suggests that organizations could invest in mentorship programs, review checklists, or dedicated training sessions to raise overall review quality. The cost of such initiatives is primarily developer time, but the payoff includes fewer production bugs and more consistent codebase standards.

Concrete examples from the author’s recent work show that human reviewers identified subtle bugs that were missed by high-end LLM reviewers, underscoring a current limitation of automated code analysis. The post also notes that existing review tools often fail to satisfy developers’ need for deep code understanding, a gap that hampers the effectiveness of both novice and experienced reviewers. Consequently, relying solely on automation may leave critical issues undiscovered.

To close the skill gap, the article proposes experimenting with new review practices and measuring their impact, though it acknowledges that human variability can be both a liability and an asset. Implementing such experiments will require allocating time for data collection and analysis, and may encounter resistance if teams are accustomed to existing workflows. The approach stops being effective where tooling cannot provide the contextual insight reviewers need, reinforcing the importance of human judgment in the review loop.